There was a similar study / survey by Microsoft (idk anymore if it was really them) recently where similar results where found. In my experience, LLM based coding assistants are pretty okay for low level complexity tasks, creating boilerplate code, especially if it does not require deeper understanding of the system architecture.

But the more complex the task becomes, the harder they start to suck and fail. This is where the time drag begins. Common mistakes or outdated coding approaches are also used rather often instead of newer standards. The deviations from the given instructions are also happening way too often. And if you do not check the generated code thoroughly, which can happen if the code “looks okay” on first glance, then finding bugs and error sources due to this can become quite cumbersome.

Debugging is where I have wasted most of my time with AI assitants. While there is some advantage in having a somewhat more capable rubber duck, it is usually not really helpful in fixing stuff. Either the error/bug sources are completely missed (even some beginner mistakes) or it tries to apply band-aid solutions rather than solving the cause or, and this is the worst of all, it is very stubborn about the alleged problem cause (possibly combined with forgetting earlier debugging findings, resulting in a tedious reasoning and chat loop). I have found myself more often than I’d like to arguing with the machine. Hallucinations or unfounded fix hypotheses make this regularly worse.
However, letting the AI assistant add some low level debug code to help analyze the problem has often been useful in my experience. But this requires clear and precise instructions, you can’t just hope the assistant will cover all important values and aspects.

When I ask the assistant to logically go through some lines of code step by step, possibly using an example, to nudge it towards seeing how it’s reasoning was wrong, it’s funny to see, e.g. with Claude, how it first says stuff like “This works as intended!” and a moment later “Wait… this is not right. Let me think about it again.”

This becomec less funny for very fundamental stuff. There were times where the AI assistant told me that 0.5 is greater than 0.8 for example, which really shows the “autocorrect on steroids” nature of LLMs rather than an active critical thinking process. This is bad, obvio. But it also makes jobs for humans in various fields of IT safe.

Typing during the whole conversation is naturally also really slow, especially when writing more than a few sentences to provide context.

Where I do find AI assistants in coding mostly useful, is in exploring APIs that I do not know so well, or code written by others that is possibly underdocumented. (Which is unfortunately really common. Most devs don’t seem to like writing documentation.)
Generating documentation for this or my own code is also pretty good most cases but also tends to contain mistakes or misses important mechanisms.

Overall in my experience I find AI assistance useful and see a mild productivity speed boost for very low level tasks with low complexity and low contextual knowledge requirements. They are useful for exploring code and writing documentation, but I can not really recommend them for debugging. It is important to learn and know how to use such AI tools precisely in order to save time instead of wasting time, since as of now they are not really capable of much.